test(daemon): de-flake sidecar stop test on the pid-liveness check

[kovtcharov-amd]

Why this matters

test_connection_stopped_after_running_raises_again fails intermittently on CI (seen on the macOS smoke runner: StopFailedError: agent 'toy' sidecar pid 4321 survived the tree-kill). It's a false failure that blocks unrelated PRs and erodes trust in the suite — the production code is fine.

Root cause: the test's fake manager reports a hardcoded pid (4321) but owns no real OS process, so its shutdown() can't make that pid disappear. registry.stop() then verifies the pid is gone via psutil.pid_exists(pid) against the real process table. On most runners pid 4321 is dead → passes; when it collides with a live process on the runner → spurious StopFailedError. Test-double fidelity bug, not a product bug: a real manager tree-kills its own real pid, so the liveness check stays meaningful in production.

Fix: patch pid_exists to False for the fake so the check reflects the double's intent (the fake sidecar is terminated) instead of the host's process table. One line + a comment; production behavior untouched.

Test plan

• pytest tests/unit/test_daemon_relay.py — 36 pass
• Verified the mechanism: forcing pid_exists → True (simulating pid 4321 alive on the runner) reproduces the StopFailedError; the fix's pid_exists → False patch neutralizes it while stop() still succeeds
• black/isort clean

[github-actions]

Verdict: Approve ✅

This de-flakes test_connection_stopped_after_running_raises_again by making the fake sidecar's shutdown believable to the liveness check. The fake reports a hardcoded pid but owns no real OS process, so stop()'s psutil.pid_exists check against the real process table would spuriously fail whenever that pid happened to be live on the runner. Patching pid_exists → False for the fake models the double's intent (the sidecar is terminated) without touching product code. Root cause and mechanism are correctly diagnosed and the fix is minimal.

🔍 Technical details

Verified

• Patch path is correct: registry.py:16 imports psutil at module scope, and registry.stop() calls psutil.pid_exists(pid) (src/gaia/daemon/sidecars/registry.py:299), so monkeypatch.setattr("gaia.daemon.sidecars.registry.psutil.pid_exists", ...) resolves to the exact symbol under test.
• Root cause matches the code: the fake manager hardcodes self.pid = 4321 (test_daemon_relay.py:224) and owns no real process, so its shutdown() can't clear that pid from the real process table — the source of the intermittent StopFailedError.
• Scope is clean: only the one flaky test changes; stop()'s liveness check is untouched, so it stays meaningful in production where a real manager tree-kills its own real pid.

No concern, noting for the record

• Forcing pid_exists → False doesn't erase coverage of the StopFailedError path — that path is exercised elsewhere (tests/integration/test_daemon_sidecars.py), and this test's assertion is that connection() raises SidecarNotRunningError after stop(), which is unrelated to the liveness branch.

Strengths

• The inline comment names the invariant (double's intent vs. host process table) in the right length — one tight block, no history dump.
• Fixes the flake at the test-double boundary rather than weakening the production assertion, which is the correct call.